Comparing early warning systems for banking crises

TL;DR: In this article, the authors assess the logit and signal extraction EWS for banking crises on a comprehensive common dataset and suggest that logit is the most appropriate approach for global EWS and signal extractor for country-specific EWS.

About: This article is published in Journal of Financial Stability.The article was published on 2008-06-01 and is currently open access. It has received 472 citations till now. The article focuses on the topics: Warning system.

Summary

1 Theoretical overview of banking crises.

• This section briefly indicates why the indicators used in EWS models such as Demirguc-Kunt and Detragiache (1998) (who use the logit method) are associated with banking crises.
• This may link to asymmetric information as if one significant bank fails, a systemic crisis may develop because the presence of such asymmetric information means depositors are unable to evaluate prospects for similar banks in terms of balance-sheet exposure to economy wide systemic risks (Kaufman and Scott, 2003).
• Accordingly, macroeconomic movements that crystallise risks particular to banking systems, namely interest rate, credit, liquidity and market risk have been the key determinants of banking crises in the last twenty years (Ergungor and Thompson, 2005).
• Alternatively, excessive market risk may be borne during boom phases, as over-optimism may concentrate portfolios in assets whose prices move procyclically, e.g. real estate (Gonzalez-Hermosillo, 1999, Craig et al 2005).
• Demirguc-Kunt and Detragiache (1998) include this variable because it indicates governments reluctance to restructure fragile banking systems and because high deficits prevent successful financial liberalisation.

2 Data, variables and specifications

• The most commonly-cited problem with EWS developed to date is the inconsistency in the banking crisis dependent variable, which is necessarily defined with a degree of subjectivity (Kaminsky and Reinhart, 1999; Demirguc-Kunt and Detragiache 1998, Eichengreen and Arteta, 2000).
• The problem lies in the fact that banking crisis is an event, so proxies for banking crises would not necessarily be perfectly correlated with banking crises themselves.
• Kaminsky and Reinhart (1999) note that crises can also be dated too early, since the worst of the crisis could unfold after the subjective start date.

2.2 The Data Sample

• The dataset of independent variables mimics the Demirguc-Kunt and Detragiache (1998) approach, utilising most of their variables but for a wider selection of countries and for a longer time span.
• The authors have used the same data source they cite: IFS and World Bank Development Indicators to obtain annual data; full data sources are obtainable in DemirgucKunt and Detragiache (1998).
• Under the DD05 dating this yields 72 systemic crisis episodes; under the less stringent CK03 definitions this yields 102 systemic crisis episodes.
• As explained in Part 1, the explanatory variables chosen are macroeconomic, financial and financial liberalisation indicators of crisis.
• 16 They find the results do not change significantly either way.

3.3 Improving the Model Further: Investigating Dynamics.

• To model the dynamics of banking crises, the authors now introduce further lags and several interaction variables so that they use the data to mimic the procyclical build up of risk.
• The authors banking crisis story unfolds as follows: credit booms are more likely to occur in an environment which allows imprudent lending, such as following the adoption of deposit insurance.
• This may be tempered if agents are not so reliant on bank intermediation for funds when deposit insurance is installed.
• The results in table 7 clearly show the procyclical behaviour of credit growth.
• From three years prior to crisis, a cyclical downturn 24 seems to occur with the coefficient sign switching so that credit rationing increases the likelihood of crisis.

3.3 In Sample Predictive Ability

• The authors now turn to see the relative performances of the different specifications in terms of their ability to accurately call crises and non-crises episodes.
• The authors initially set their threshold much higher than Demirguc-Kunt and Detragiache (1998) who set their cut off probability at 0.0519.
• In contrast the authors set their threshold at 0.5, arguing that from the policy maker s perspective, costly intervention on the basis of a crude EWS should be avoided unless the model seriously calls a crisis (table 8a).
• Their models out-perform the DD05 models in terms of crisis prediction (apart from regressions 1 and 2).the authors.
• This may be due to the wider definitions of crisis that CK03 use so that the CK03 dummy is associated with a reduced ability to call non-crisis periods correctly and therefore a higher chance of Type II errors.

3.5 Furthering the Signal Extraction Approach: Composite Indicators

• Borio and Lothe authors (2002) developed the Kaminsky and Reinhart (1999) procedure by constructing composite indicators22 to extract signals of banking crisis.
• In their case, the authors set stronger criteria; they first construct a composite indicator using variables that have a noise to signal ratio of 0.5 or less and then they construct composites using variables where the ratio is 0.75 or less.
• For the 0.75 cut-off the authors construct two composites based on the general and country specific approaches.
• The composite method therefore appears to improve crisis prediction whether a country specific or general threshold approach is used but this occurs at the cost of higher Type II errors.
• Conversely, the specific threshold approach which already yields low Type I errors benefits from the composite method so that the composite specific threshold approach lowers the NTSR below that of virtually all the univariate indicators with the same methodology.

3.6 Comparison of the Signal Extraction and Multivariate Logit Models.

• On the basis of in-sample predictive ability, the multivariate logit model outperforms the signal extraction approach in terms of the percentage of crises correctly predicted.
• With the signal extraction approach, every indicator misses a substantial number of crises.
• If the NTSR jumps at different threshold for different countries, then using cross-country data to derive a common threshold may not be fruitful; the threshold that minimises the NTSR will fall on either side of the individual optimal thresholds for most countries.
• An alternative approach to ours would be the Borio and Lothe authors (2002) methodology where individual variable thresholds are jointly evaluated.
• Composite indicator ability to predict crises is more comparable to the multivariate logit model than the univariate signal extraction procedure so the two procedures could be compared on the same dataset.

4 Conclusion

• A comparison of the multinomial logit and signal extraction procedures shows that real GDP growth and terms of trade are robust leading indicators of banking crisis for their comprehensive sample.
• Creating composite indicators may further improve crisis prediction.
• Systemic Banking Crises , Policy Discussion Papers, Federal Reserve Bank of Cleveland, Number 9, February 2005.
• Finance and Development, Volume 36, Number 2, IMF Publications, Washington.

Figures

Table 5: Regression 4: All Variables

Table 8a: In Sample Predictive Ability: Cut-Off Probability = 0.5

Table 8a: In Sample Predictive Ability: Cut-Off Probability = 0.5

Table 6: Regression 5: All Variables: Standardisation, More Lags and Logs Introduced

Table 10: Variables Qualifying for the Composite Indicator

Table 4: Regression 3: Macroeconomic, Financial and Fiscal Variables.

Table 11: Comparison of In Sample Predictive Ability of Composite Indicators Using General and Country Specific Thresholds.

Table 3: Regression 2: Macroeconomic (minus fiscal) and Financial Variables

Table 9: Comparison of In-Sample Predictive Ability for General and Country Specific Thresholds.

Table 1: The Demirguc-Kunt and Detragiache (1998) Variables

Table 7: Regression 6: Procyclicality and Moral Hazard

Table 2: Regression 1: Macroeconomic (minus fiscal) Variables Only

Frequently asked questions

Q1. What have the authors contributed in "Comparing early warning systems for banking crisis" ?

In this context, the authors assess the logit and signal extraction EWS for banking crises on a comprehensive common dataset. The authors suggest that logit is the most appropriate approach for global EWS and signal extraction for country specific EWS. Furthermore it is important to consider the policy maker’s objectives when designing predictive models and setting related thresholds since there is a sharp trade-off between correctly calling crises and false alarms. 

Q2. What have the authors stated for future works in "Comparing early warning systems for banking crisis" ?

4. Since these two approaches have been shown to favour Type I and Type II errors differently, the authors wish to see if this behaviour persists when they construct composites. 

Q3. What are the main factors that could trigger cyclical downturns?

Macroeconomic shocks which could trigger cyclical downturns thereby increasing NPLs include adverse movements in terms of trade and correspondingly currency depreciations, especially for small open economies. 

Q4. What are the key determinants of banking crises?

macroeconomic movements that crystallise risks particular to banking systems, namely interest rate, credit, liquidity and market risk have been the key determinants of banking crises in the last twenty years (Ergungor and Thompson, 2005). 

Q5. Why is asymmetric information important for loan officers?

Because asymmetric information becomes disproportionately important for loan officers, lending spreads are artificially high and intermediaries hold excess capital and provisions. 

Q6. What is the effect of financial liberalisation on credit risk?

In liberalised markets, increased competition may erode bank charter values so that without adequate supervision and regulation, banks forgo prudent credit risk assessment in a bid to catch borrowers. 

Q7. How are real interest rates able to call crises correctly?

For Kaminsky and Reinhart (1999), real interest rates are able to call 100% of crises correctly whilst domestic credit/ GDP is able to call 50% of crises correctly. 

Q8. What is the reason for the high levels of NPLs associated with crises?

Although banks enjoy advantages in screening and monitoring borrowers, both of which reduce credit risk, the high levels of NPLs associated with crises indicate risk assessment by banks deteriorates during pre-crisis periods. 

Q9. Why does asymmetric information restrict credit availability?

Asymmetric information does not restrict credit availability because bank managers succumb to euphoric and herding behaviour7, utilising biased information sets to make investment decisions. 

Q10. What criteria are used to determine the extent of non-performing loans?

They stipulate that non-performing loans as a proportion of entire loans of the banking system must be in the range of 5 10% or less. 

Q11. What is the effect of low-cost deposit financing on bank balance sheets?

during booms, banks may use low-cost deposit financing to invest heavily in particular sectors which appear profitable and where collateral values are high. 

Q12. What is the likely scenario for a credit boom?

Their banking crisis story unfolds as follows: credit booms are more likely to occur in an environment which allows imprudent lending, such as following the adoption of deposit insurance. 

Q13. Why do Berg and Pattillo (1999) suggest that the signal extraction process is not?

Berg and Pattillo (1999) suggest this is because the signal extraction process assumes that a threshold for a variable is a discrete value and that whenever this is crossed, a crisis becomes impending. 

Q14. What does the fact that deposit insurance is installed in the banking crisis story show?

The fact that the interaction terms are significant (with credit growth alone insignificant) also demonstrates how moral hazard is more prevalent when deposit insurance exists so that credit booms generate considerable banking crisis risk. 

Q15. What is the effect of deposit insurance on the likelihood of a crisis?

The safety net of deposit insurance significantly raises the likelihood of morally hazardous lending by banks, which adds to the crisis probability. 

Q16. What is the link between credit risk and market risk?

Since credit risk is also procyclical, the link between the two risks becomes apparent when asset price collapses realise market risk and low collateral values realise credit risk.